Improving ocean color algorithms for Chlorophyll-a retrieval with Machine learning and ensemble modeling in optically complex coastal waters

Titelangaben

Moradi, Masoud ; Lu, Meng ; van der Molen, Johan ; Arabi, Behnaz:
Improving ocean color algorithms for Chlorophyll-a retrieval with Machine learning and ensemble modeling in optically complex coastal waters.
In: ISPRS Journal of Photogrammetry and Remote Sensing. Bd. 236 (2026) . - S. 212-238.
ISSN 0924-2716
DOI: https://doi.org/10.1016/j.isprsjprs.2026.03.045

Volltext

Link zum Volltext (externe URL):

Abstract

Accurate retrieval of Chlorophyll-a concentrations (Chla, mg m−3) in optically complex and turbid coastal waters remains challenging due to overlapping signals from other water constituents, which obscure the phytoplankton reflectance. Machine Learning (ML) models trained directly on raw spectral bands or simple band combinations may learn sensor- or region-specific artifacts rather than the true biophysical relationship. This limits their transferability across sensors, regions, and seasons, and increases the risk of overfitting when applied to new environment. In this study, we developed and validated a hybrid approach that integrates optimized conventional ocean color algorithms with ML techniques to estimate Chla from Sentinel-2 and Sentinel-3 imagery in the Dutch Wadden Sea, the Netherlands, a highly turbid meso-tidal area. First, empirical and semi-analytical ocean color algorithms were spectrally tuned via a genetic algorithm against in-situ matchups of remote sensing reflectance (Rrs(λ), sr-1) and measured Chla. Significant spectral features of Rrs(λ) spectra are further identified through first- and second-derivative analysis. The outputs of these optimized algorithms, along with featured reflectance bands, were then used as inputs to six prominent ML models. Shapley Additive Explanations (SHAP) method was applied to identify the most informative input variables, leading to the selection of five algorithmic indices and four spectral bands. An ensemble learning strategy, specifically stacking with ridge regression, was employed to combine the predictions of the best-performing models into a final model. The ensemble ML model achieved the highest accuracy, with R2 = 0.92 and RMSE = 2.55 mg m−3 for in-situ matchups, and R2 = 0.73 − 0.79 and RMSE = 2.75 − 2.71 mg m−3 for Sentinel-2 − Sentinel-3 imagery matchups, respectively. To assess its transferability, the ensemble ML model was applied in Chesapeake Bay (East Coast of the USA), where it produced reasonable spatial Chla patterns and maintained good performance from Sentinel-2 and Sentinel-3 images (R2 values of 0.75 and 0.81, and RMSE of 2.9 and 2.3 mg m−3, respectively). Overall, the results demonstrate that embedding physically interpretable optimized ocean color algorithms within a ML data-driven framework enhances robustness, reduces overfitting, and improves transferability of Chla retrievals in optically complex coastal waters.